WanAnimatePlus ModelLoader

Every WanAnimatePlus workflow starts here. ModelLoader takes your diffusion checkpoint out of ComfyUI/models/diffusion_models, applies precision, quantization, attention and memory settings, optionally bolts on a LoRA, VACE, or one of the talking/portrait extra models, and hands the chain a single WANVIDEOMODEL. Get this node wrong and nothing downstream matters - so here's what each family of settings actually does.

Precision and quantization. base_precision defaults to bf16, which is right for most modern cards. quantization defaults to disabled - and the tooltip explains that "disabled" really means "autoselect based on weights," so it's not as lazy as it sounds. The fp8 modes are where the VRAM wins live: fp8_e4m3fn and friends. Watch the compute-capability note though - the _fast modes (fp8 matmul) need CUDA compute capability 8.9+ (RTX 4000 series and up), and e4m3fn generally can't be torch.compiled on 3000-series and older. If you're on a 3090, that's the difference between a working fp8 load and a crash.

Where the model lives. load_device defaults to offload_device - the model starts in your system RAM and blocks shuffle in as needed. The tooltip is a warning in disguise: "NOT recommended with the larger models unless you have 48GB+ VRAM" if you set it to main_device. Leave the default; that's the whole point of block swap and VRAM management.

Attention and speed. attention_mode (default sdpa) is your menu: flash_attn_2, sageattn, sageattn_compiled, radial_sage_attention, and so on. SageAttention is the community's favorite speedup but is famously fiddly to install on Windows (Triton). compile_args hooks in torch.compile via the pack's TorchCompileSettings; rms_norm_function (pytorch is the newer native RMSNorm, faster without compile, slightly different results).

The memory helpers. block_swap_args takes the output of WanAnimatePlus BlockSwap - the CPU-RAM overflow parking for small cards. vram_management_args is an alternative from DiffSynth-Studio: more aggressive memory reduction than block swap, but slower. You generally pick one, not both. lora takes your WANVIDLORA chain from LoraSelect/LoraSelectMulti, and extra_model accepts a VACE or MTV Crafter to stack on top.

Install via ComfyUI Manager (search "WanAnimatePlus") or git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus into custom_nodes, restart, and put your checkpoint in ComfyUI/models/diffusion_models - that folder is the only place this node looks. The fork's README is explicit that the original ComfyUI-WanVideoWrapper should be installed alongside, and that you must use the WanAnimatePlus chain end to end. If your load fails on a 3000-series card with a fp8 mode, the compute-capability note above is your answer.